Mitigating Overflow of Object Detection Tasks Based on Masking Semantic Difference Region of Vision Snapshot for High Efficiency
Heuijee Yun, Daejin Park · 2022
Object recognition functions are essential to properly perform safety and autonomous driving functions. However, sophisticated object recognition work requires extensive computation. It is difficult to handle a large amount of computation on the lightweight embedded boards currently used in vehicles. In this paper, we propose a method using machine learning and deep learning for lightweight object recognition algorithm in lightweight embedded boards. We created an algorithm suitable for lightweight embedded boards by appropriately using deep neural network architecture that requires small computational volumes but provides low accuracy, as well as deep-learning algorithms that require large computational volumes but provide high accuracy. After determining the area using a deep neural network architecture algorithm with a relatively small amount of computation, we improved the accuracy by using a more accurate deep learning algorithm. We used OpenCV to process input images in Python, and we processed image by using efficient neural network (ENet) and You Only Look Once (YOLO). By executing this algorithm, we can realize more accurate and lightweighted object recognition.